rio

One package, four submodules:

• rio.ai -- a multi-provider LLM streaming SDK (Anthropic, Google Gemini, Mistral, OpenAI Codex, OpenAI-compatible), ported from huggingface/tau's tau_ai module. The message/tool/type vocabulary tau_ai depends on (tau_agent.messages, tau_agent.tools, tau_agent.types, tau_agent.provider, tau_agent.provider_events upstream) is folded in here too, so rio.ai is self-contained. See NOTICE for the upstream MIT license.
• rio.agent -- a from-scratch runtime implementing SKILL.state: Scalable Long-Horizon Agent Skills (Badhe, Tiwari & Chung; EMNLP), built on rio.ai.
• rio.coding -- the coding backend (CLI, tools, skills, extensions), built on rio.agent.
• rio.tui -- the terminal application, wired to rio.coding. It owns session tabs, the prompt editor, shell, file navigation, diffs, and settings.

Coding agent

uv sync
uv run rio setup --provider local --base-url http://localhost:8080/v1 --model my-model
uv run rio                         # interactive terminal interface
uv run rio -p "Inspect this project and explain its entry point"
uv run rio -p --mode json "Add a regression test for the parser"
uv run rio --mode rpc              # JSONL commands on stdin
uv run rio sessions
uv run rio --session SESSION_ID
uv run rio export SESSION_ID --format html

Choose an existing provider with --provider NAME --model MODEL; rio providers lists the configured catalog. Credentials may come from the provider's environment variable or Rio's credential store (rio login PROVIDER; use --method api-key for an API key). A successful login remembers the provider for future launches. API keys and OAuth tokens are stored in ~/.rio/credentials.json with owner-only permissions; the file is unencrypted. OAuth tokens refresh automatically. Use rio logout PROVIDER to remove saved credentials. Configuration and journals live under ~/.rio. Use --approve to allow ambient project instructions and extensions for a run; project trust controls resource loading, not what shell commands can access.

The coding tools are read, write, edit, bash, and respond. Each step updates structured state and executes one action. Sessions store snapshots for resume and checkpoint restoration; they never replay a transcript into the model. HTML exports show steps, final state, and estimated token footprints.

See the installed documentation for providers, skills, extensions, RPC, and the state-based runtime. The offline example runs a complete session with a fake provider and requires no credentials.

Development

uv sync
uv run pytest
uv run ruff check .

Tests use scripted providers (rio.ai.FakeProvider) and mocked HTTP; no live provider credentials are required. Headless Textual tests exercise terminal interaction, and tool integration tests work in temporary directories. See in particular tests/test_rio_agent_loop.py::test_prompt_footprint_is_bounded_across_steps, which asserts that per-step prompt size stays constant across many steps rather than growing -- a direct runtime check of the paper's core claim.